SPECfp

SPECfp is a floating-point benchmark suite developed by the Standard Performance Evaluation Corporation (SPEC) to measure the floating-point performance of computer systems. It consists of multiple programs written in Fortran and C that exercise a range of floating-point operations, from linear algebra to numerical simulations, to reflect real-world FP workloads.

Several revisions of SPECfp have been released, including SPECfp92, SPECfp95, and SPECfp2000, with later alignment under the SPEC CPU family (such as CPU2006 and CPU2017). Each version updates its workloads and input data to keep pace with contemporary hardware and compiler technology. The results are typically reported as a SPECfp score, with higher values indicating better performance; some variants also report throughput scores like SPECfp_rate.

Testing is performed by compiling the benchmarks with standardized settings, executing them on the target system, and recording execution times. The overall score is derived by normalizing each program's time against a reference machine and taking a geometric mean across the suite. Some releases separate base scores (fixed settings) from peak or rate scores to improve comparability.

SPECfp results are used by hardware vendors, researchers, and evaluators to compare floating-point performance across architectures and compilers. Limitations include that the workloads may not cover all FP tasks, and results depend on software and environmental conditions; therefore, scores should be interpreted in a controlled, reproducible context.